Local Open TTS Models: VRAM and RAM Requirements, Measured
In short: Most open-source local TTS models run in 4GB of VRAM or less, and several — Kokoro, Piper, and MeloTTS — synthesize speech in real time on CPU with no GPU at all. Only large voice-cloning models are heavy: Tortoise and Sesame CSM want about 8GB. So the common belief that you need a strong GPU for text-to-speech is mostly wrong.
Most open-source local TTS models run in 4GB of VRAM or less, and several — Kokoro, Piper, and MeloTTS — synthesize speech in real time on CPU with no GPU at all. Only large voice-cloning models are heavy: Tortoise and Sesame CSM want about 8GB. So the common belief that you need a strong GPU for text-to-speech is mostly wrong.
In one line: for narration, CPU is enough (Kokoro, Piper); for voice cloning, 4-6GB VRAM (XTTS-v2, F5-TTS); only the largest, most natural models need 8GB+.
What decides TTS VRAM?#
Unlike an LLM, TTS memory is not driven by context (a KV cache) but by model parameters plus the vocoder and sample rate. An acoustic model turns text into audio features, and a vocoder (for example HiFi-GAN) turns those into a waveform; both sit in memory together. Longer text is processed in chunks, so memory does not balloon — which is why most TTS models are lighter than an 8B LLM (about 5GB at 4-bit).
| Model | Params | VRAM | CPU/RAM only | Voice clone | License |
|---|---|---|---|---|---|
| Kokoro-82M | 82M | ~1-3GB | Yes (~10x real-time) | No | Apache-2.0 (commercial OK) |
| Piper | Small | None (CPU) | Yes (Raspberry Pi) | No | MIT -> GPL-3.0 (fork) |
| MeloTTS | Small | ~1-2GB | Yes | No | MIT (commercial OK) |
| XTTS-v2 | ~0.5B | ~4-6GB | Limited (16GB Mac RAM) | Yes (17 langs, 6s) | CPML (non-commercial) |
| F5-TTS | ~0.3B | ~2-4GB | Limited | Yes (3s) | CC-BY-NC (non-commercial) |
| Fish Speech 1.5 | Large DualAR | ~4GB | Limited | Yes (strong CJK) | Varies by version |
| Tortoise / Sesame CSM | Large | ~8GB | Hard | Yes | Check each repo |
Can it really run on CPU only?#
Yes. Kokoro-82M reaches roughly 10x real-time on CPU and fits in 1-3GB VRAM on almost any integrated or older GPU. Piper needs no GPU at all and runs in real time on embedded devices like a Raspberry Pi (30+ languages). For narration, caption voice-over, or notification speech — anything that does not need cloning — a CPU model is the right answer: negligible power, heat, and cost.
How much for voice cloning?#
Cloning models add a reference-audio encoder, so they use a bit more. XTTS-v2 clones into 17 languages from a 6-second sample in about 4-6GB and runs well on Apple Silicon unified memory (16GB RAM). F5-TTS uses a flow-matching design, clones from 3 seconds in ~2-4GB, and is faster at 5-8x real-time on an RTX 4070. Both are non-commercial licensed (and XTTS has no one left to sell a commercial license after Coqui shut down), so verify before shipping.
Which should you pick for your hardware?#
No GPU or under 4GB: Kokoro or Piper (both commercial-friendly). An 8-12GB GPU: F5-TTS or Fish for cloning (check the license). A unified-memory Mac uses RAM as VRAM, so XTTS-v2 is fine too. Only reach for the largest models if you need maximum naturalness and have 8GB+ — most real work fits under 4GB.
Note: figures are 2025-2026 community-measured ranges (fp16; they vary with quantization, batch, and sample rate) and licenses differ per release, so check each model card before deploying. Open TTS moves fast, so this is reviewed quarterly.
Related reading: TTS도 이종교배가 될까 — 코덱·공통조상·채점 3중고, 로컬 오픈 LLM VRAM·RAM 요구량, 직접 계산·실측
References#
- Kokoro-82M (Hugging Face) — 82M, Apache-2.0, CPU-capable
- Piper (OHF-Voice fork) — CPU-only, 30+ languages (old rhasspy/piper archived 2025-10)
- F5-TTS (GitHub) — flow-matching, 3s cloning
- Coqui XTTS-v2 (GitHub) — 17-language cloning, CPML non-commercial
- Fish Speech (GitHub) — DualAR, strong CJK
Responses
No responses yet. Be the first to respond.